New Asymmetric Test Detects Distribution Drifts in LLM Outputs

This paper introduces ASK-NN, an Asymmetric Nearest-Neighbor test specifically designed to identify distribution drifts in natural language data. The motivation stems from the common problem of detecting "hallucinations" or artificially generated text in large language model (LLM) outputs, where the statistical distribution of the response might deviate from that of the input prompt or retrieved context. Unlike traditional two-sample tests, ASK-NN is asymmetric, meaning it treats the reference (e.g., prompt) and query (e.g., response) samples differently, which is crucial given their often significant length disparities. The test statistic counts reference points whose nearest neighbor in the combined sample is also a reference point. The authors provide theoretical guarantees for its statistical properties, including asymptotic normality. Empirical evaluations demonstrate that ASK-NN is computationally efficient and performs competitively against other kernel and graph-based baseline methods. It shows promise in synthetic benchmarks, artificial text detection, and, importantly, in identifying LLM hallucinations by analyzing token-level hidden states.